pith. sign in

arxiv: 2606.21828 · v1 · pith:RC2KUCIAnew · submitted 2026-06-20 · 🧮 math.NA · cs.LG· cs.NA

Spectrally Safe Neural Operator Warm-Starts for Large-Scale Newton Solvers

Jaemin Oh , Youngkyu Lee , Jerome Darbon , George Em Karniadakis This is my paper

Pith reviewed 2026-06-26 12:08 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NA
keywords neural operatorsNewton solvershyperelasticityJacobian spectrumwarm-startenergy penalizationlarge-scale PDEindefinite Jacobian
0
0 comments X

The pith

Neural operators with O(10^{-3}) L2 error can still produce indefinite Jacobians for Newton solvers, but energy penalization fixes the spectrum without new data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates that low relative L2 error in a neural operator does not ensure the predicted field lies in the basin of attraction for Newton solvers of nonlinear PDEs, because mean-squared training overlooks localized pointwise violations of constraints such as volume preservation. In a nearly incompressible hyperelasticity problem, this causes det F to deviate from one, producing negative eigenvalues in the discrete Jacobian even when the field appears accurate. A short label-free fine-tuning phase that adds a penalty on the discrete energy shifts the Jacobian spectrum back to positive definite. Combined with an inexact outer loop, the resulting warm-started Newton method then converges across the full loading range on problems with millions of degrees of freedom.

Core claim

An operator trained to relative L2 error O(10^{-3}) can produce initial states in which det F is dispersed away from one, so that the discrete Jacobian acquires negative eigenvalues even when the predicted field is visually close to the reference. Penalizing the operator against the discrete energy in a short label-free fine-tuning phase restores positive definiteness, allowing inexact Newton iterations to succeed across the full loading range where the unregularized operator fails and yielding up to 5.4 times wall-clock speedup on a 3D problem with 6.4 million degrees of freedom.

What carries the argument

The label-free fine-tuning phase that penalizes the neural operator output against the discrete energy functional to enforce positive definiteness of the Jacobian.

If this is right

  • The combined warm-started Newton method converges across the full loading range where the unregularized operator fails.
  • It reaches up to 5.4 times wall-clock speedup over incremental continuation on a 3D hyperelasticity problem with 6.4 million degrees of freedom.
  • No additional labeled solution data are needed for the fine-tuning step.
  • The approach remains compatible with inexact outer loops and memory-feasible Krylov solvers at multi-million-dof scale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same energy penalty may stabilize neural warm-starts for other nonlinear PDEs whose Jacobians become indefinite when physical invariants are locally violated.
  • Reducing dependence on incremental continuation could simplify simulation workflows for large-scale nonlinear mechanics.
  • Testing the fine-tuned operator on refined meshes or altered material parameters would check whether spectral safety holds under changes in discretization.

Load-bearing premise

Penalizing the operator against the discrete energy during a short label-free fine-tuning phase will reliably shift the Jacobian spectrum to positive definite across the full loading range without requiring additional solution data or introducing new instabilities.

What would settle it

Apply the fine-tuned operator to the full loading path on the 6.4-million-dof mesh and check whether every resulting discrete Jacobian has strictly positive eigenvalues and whether Newton iterations converge at every load step.

Figures

Figures reproduced from arXiv: 2606.21828 by George Em Karniadakis, Jaemin Oh, Jerome Darbon, Youngkyu Lee.

Figure 1
Figure 1. Figure 1: Numerical behavior and optimization trajectories for the pedagogical toy [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Numerical discretization and steady-state solution profiles for the lid-driven [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Residual norm convergence profiles for hybrid and classical numerical solvers [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overlay of the original domain and deformed domain for [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The 20 smallest eigenvalues of the initial system Jacobian matrices evaluated [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spatial distributions and statistical frequencies of the deformation gradient [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Convergence plots of the relative residual norm [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Quantitative performance profiles across a parameter sweep [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
read the original abstract

Neural operators are increasingly used to warm-start Newton solvers for nonlinear PDEs, on the premise that a low test error places the initial guess inside the basin of attraction. We show that this premise is unreliable. An operator trained to the relative \(L^2\) error \(O(10^{-3})\) can still produce an initial state in which the discrete Jacobian is indefinite, because the mean-squared training controls error on average while leaving localized pointwise violations of the underlying physics. For a nearly incompressible hyperelasticity problem, we trace this to the predicted volume change: the operator disperses \(\mathrm{det} F\) well away from one, and the resulting Jacobian acquires negative eigenvalues even when the predicted field is visually indistinguishable from the reference. At a small scale, this is a nuisance; at a multi-million degree-of-freedom scale, it is disqualifying, since the conjugate gradient and other Krylov solvers needed for memory-feasible Newton steps assume a definite spectrum. We then show that a short, label-free fine-tuning phase -- penalizing the operator against the discrete energy, with no additional solution data -- shifts the Jacobian spectrum back to positive definite. Combined with an inexact outer loop, this gives a warm-started Newton method that converges across the full loading range where the unregularized operator fails, reaching up to 5.4\(\times\) wall-clock speedup over incremental continuation on a 3D problem with 6.4 million degrees of freedom.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript argues that neural operators trained only to relative L2 error O(10^{-3}) are unreliable warm-starts for Newton solvers on nonlinear PDEs because mean-squared loss permits localized pointwise violations of the physics (e.g., det F far from 1 in hyperelasticity), producing indefinite discrete Jacobians even when the predicted field looks plausible. It then introduces a short label-free fine-tuning stage that penalizes the operator output against the discrete energy, shifting the Jacobian spectrum to positive definite and enabling convergence over the full loading range; combined with an inexact outer loop this yields up to 5.4× wall-clock speedup versus incremental continuation on a 6.4-million-DOF 3D hyperelasticity problem.

Significance. If the energy-penalized fine-tuning reliably restores positive-definiteness across the loading path without degrading accuracy or creating new instabilities, the work removes a practical obstacle to deploying neural operators inside memory-constrained large-scale nonlinear solvers that must rely on Krylov methods. The concrete demonstration on a multi-million-DOF example supplies a falsifiable performance claim that can be checked by other groups.

major comments (2)
  1. [Methods / fine-tuning description] The central claim that the label-free energy penalty corrects localized det F violations and produces positive-definite Jacobians for every load step rests on an unstated penalty formulation and weight; without the explicit expression (presumably in the methods section) or an ablation on the weight, it is impossible to verify that the correction targets the physics violation rather than driving the operator to a different local minimum.
  2. [Results / large-scale experiment] The reported 5.4× speedup and full-range convergence are load-bearing for the practical contribution, yet the abstract supplies neither error bars on the timing, nor success/failure rates across the loading path, nor a comparison against the unregularized operator on the same 6.4 M DOF mesh; these omissions leave the weakest assumption (that the fine-tuning never introduces new instabilities) untested.
minor comments (2)
  1. The abstract states that the operator is trained to relative L2 error O(10^{-3}) but does not report the precise test-set error or the training/validation split used for the hyperelasticity example.
  2. A plot of the smallest Jacobian eigenvalue versus load step, before and after fine-tuning, would make the spectral correction claim immediately verifiable.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and for recognizing the potential impact on memory-constrained large-scale nonlinear solvers. We respond to each major comment below and indicate the changes that will be incorporated in the revised manuscript.

read point-by-point responses

  1. Referee: [Methods / fine-tuning description] The central claim that the label-free energy penalty corrects localized det F violations and produces positive-definite Jacobians for every load step rests on an unstated penalty formulation and weight; without the explicit expression (presumably in the methods section) or an ablation on the weight, it is impossible to verify that the correction targets the physics violation rather than driving the operator to a different local minimum.

    Authors: We agree that the explicit mathematical expression for the energy penalty and the numerical value of the weight were not stated with sufficient precision in the methods section. This was an oversight that hinders reproducibility. In the revised manuscript we will insert the precise formulation of the label-free penalty (the discrete hyperelastic energy evaluated directly on the operator output) together with the specific weight chosen on a small validation set. We will also add a short ablation table showing the effect of varying the weight on both L2 accuracy and the fraction of load steps for which the Jacobian remains positive definite. revision: yes

  2. Referee: [Results / large-scale experiment] The reported 5.4× speedup and full-range convergence are load-bearing for the practical contribution, yet the abstract supplies neither error bars on the timing, nor success/failure rates across the loading path, nor a comparison against the unregularized operator on the same 6.4 M DOF mesh; these omissions leave the weakest assumption (that the fine-tuning never introduces new instabilities) untested.

    Authors: The abstract is length-constrained and therefore omits these statistics, but the main text already documents that the unregularized operator fails to converge on the full loading path while the fine-tuned operator succeeds, together with the observed wall-clock speedup versus incremental continuation. Nevertheless, we accept that explicit variability measures, per-step success rates, and a direct side-by-side comparison on the identical 6.4 M DOF mesh would make the claim more robust. In the revision we will add timing statistics with standard deviations, a table of convergence success/failure for each load increment, and a column-wise comparison against the unregularized operator on the same mesh. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; fine-tuning presented as independent adjustment

full rationale

The paper identifies that L2 training to O(10^{-3}) error can leave localized det F violations producing indefinite Jacobians, then introduces a separate label-free energy-penalization fine-tuning phase as a post-training correction. No equation reduces the spectral correction, convergence across loading range, or reported 5.4x speedup to a quantity defined by the original training loss or fitted parameters. The demonstration on the 6.4M-DOF problem is empirical rather than a self-referential derivation, and no self-citation chain is invoked to justify uniqueness or the penalty form. This yields a self-contained empirical claim with at most minor non-load-bearing self-citation risk.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The abstract supplies limited detail; the method introduces at least one tunable penalty weight and relies on the standard mathematical requirement that Krylov solvers need positive-definite Jacobians.

free parameters (1)
  • energy penalty weight
    The coefficient controlling the strength of the discrete-energy penalty in the fine-tuning objective is a free parameter whose value must be chosen or tuned.
axioms (1)
  • standard math Conjugate gradient and other Krylov solvers require a positive-definite Jacobian to converge reliably.
    Invoked when stating that an indefinite spectrum disqualifies the warm-start for memory-feasible Newton steps at large scale.

pith-pipeline@v0.9.1-grok · 5809 in / 1456 out tokens · 43198 ms · 2026-06-26T12:08:07.002141+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

270 extracted references · 6 canonical work pages

  1. [1]

    Mathematics of Computation , pages=

    Simple Finite Element Method in Vorticity Formulation for Incompressible Flows , author=. Mathematics of Computation , pages=. 2001 , publisher=

    2001

  2. [2]

    SIAM review , volume=

    Iterative methods by space decomposition and subspace correction , author=. SIAM review , volume=. 1992 , publisher=

    1992

  3. [3]

    International journal for numerical methods in engineering , volume=

    Gmsh: A 3-D finite element mesh generator with built-in pre-and post-processing facilities , author=. International journal for numerical methods in engineering , volume=. 2009 , publisher=

    2009

  4. [4]

    Hunter, J. D. , Title =. Computing in Science & Engineering , Volume =

  5. [5]

    Proceedings of the National Academy of Sciences , volume=

    Elliptic PDE learning is provably data-efficient , author=. Proceedings of the National Academy of Sciences , volume=. 2023 , publisher=

    2023

  6. [6]

    IMA Journal of Numerical Analysis , pages=

    Error analysis for finite element operator learning methods for solving parametric second-order elliptic PDEs , author=. IMA Journal of Numerical Analysis , pages=. 2026 , publisher=

    2026

  7. [7]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Spectral operator learning for parametric PDEs without data reliance , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2024 , publisher=

    2024

  8. [8]

    SIAM Journal on Scientific Computing , volume=

    Finite element operator network for solving elliptic-type parametric pdes , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

    2025

  9. [9]

    SIAM Journal on Scientific Computing , volume=

    Nonadaptive quasi-optimal points selection for least squares linear regression , author=. SIAM Journal on Scientific Computing , volume=. 2016 , publisher=

    2016

  10. [10]

    Spectral/hp Element Methods for Computational Fluid Dynamics , year =

    Karniadakis, George and Sherwin, Spencer , doi =. Spectral/hp Element Methods for Computational Fluid Dynamics , year =

  11. [11]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Squeeze-and-excitation networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  12. [12]

    2003 , publisher=

    Introduction to numerical continuation methods , author=. 2003 , publisher=

    2003

  13. [13]

    2004 , publisher=

    Domain decomposition methods-algorithms and theory , author=. 2004 , publisher=

    2004

  14. [14]

    International Conference on Machine Learning , pages=

    Learning to optimize multigrid PDE solvers , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  15. [15]

    International journal for numerical methods in fluids , volume=

    Discussions on driven cavity flow , author=. International journal for numerical methods in fluids , volume=. 2009 , publisher=

    2009

  16. [16]

    Annual review of fluid mechanics , volume=

    Fluid mechanics in the driven cavity , author=. Annual review of fluid mechanics , volume=. 2000 , publisher=

    2000

  17. [17]

    Journal of computational physics , volume=

    High-Re solutions for incompressible flow using the Navier-Stokes equations and a multigrid method , author=. Journal of computational physics , volume=. 1982 , publisher=

    1982

  18. [18]

    SIAM Journal on Matrix Analysis and Applications , volume=

    A fully asynchronous multifrontal solver using distributed dynamic scheduling , author=. SIAM Journal on Matrix Analysis and Applications , volume=. 2001 , publisher=

    2001

  19. [19]

    SIAM Journal on Scientific Computing , volume=

    On the training and generalization of deep operator networks , author=. SIAM Journal on Scientific Computing , volume=. 2024 , publisher=

    2024

  20. [20]

    arXiv preprint arXiv:2511.19980 , year=

    Operator learning at machine precision , author=. arXiv preprint arXiv:2511.19980 , year=

    arXiv

  21. [21]

    SIAM Journal on Scientific Computing , volume=

    Deeponet based preconditioning strategies for solving parametric linear systems of equations , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

    2025

  22. [22]

    International Conference on Machine Learning , pages=

    Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  23. [23]

    SIAM Journal on Scientific Computing , volume=

    Choosing the forcing terms in an inexact Newton method , author=. SIAM Journal on Scientific Computing , volume=. 1996 , publisher=

    1996

  24. [24]

    Technical Report ANL-21/39 - Revision 3.24

    Balay, S. and Abhyankar, S. and Adams, M. F. and Brown, J. and Brune, P. and Buschelman, K. and Constantinescu, E. M. and Dalcin, L. and Benson, S. and Dener, A. and others , title =. 2025 , month =. doi:10.2172/2998643 , url =

    work page doi:10.2172/2998643 2025

  25. [25]

    Barrata, Joseph P

    Baratta, Igor A. and Dean, Joseph P. and Dokken, J. 2023 , bdsk-url-1 =. doi:10.5281/zenodo.10447666 , howpublished =

    work page doi:10.5281/zenodo.10447666 2023

  26. [26]

    Alnæs, Anders Logg, Kristian B

    Alnaes, Martin S. and Logg, Anders and. Unified Form Language: A domain-specific language for weak formulations of partial differential equations , volume =. 2014 , bdsk-url-1 =. doi:10.1145/2566630 , journal =

    work page doi:10.1145/2566630 2014

  27. [27]

    SIAM Journal on Scientific Computing , volume=

    An analysis of approximate nonlinear elimination , author=. SIAM Journal on Scientific Computing , volume=. 1996 , publisher=

    1996

  28. [28]

    International conference on machine learning , pages=

    Spherical fourier neural operators: Learning stable dynamics on the sphere , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  29. [29]

    Physical Review Research , volume=

    Multifidelity deep neural operators for efficient learning of partial differential equations with application to fast inverse design of nanoscale heat transport , author=. Physical Review Research , volume=. 2022 , publisher=

    2022

  30. [30]

    Transolver: A Fast Transformer Solver for

    Haixu Wu and Huakun Luo and Haowen Wang and Jianmin Wang and Mingsheng Long , booktitle=. Transolver: A Fast Transformer Solver for. 2024 , url=

    2024

  31. [31]

    International Conference on Learning Representations , year=

    Fourier Neural Operator for Parametric Partial Differential Equations , author=. International Conference on Learning Representations , year=

  32. [32]

    Uspekhi Mat Nauk , volume=

    Functional analysis and applied mathematics (in Russian) , author=. Uspekhi Mat Nauk , volume=

  33. [33]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Nows: Neural operator warm starts for accelerating iterative solvers , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2026 , publisher=

    2026

  34. [34]

    Journal of Computational Physics , volume=

    Neural operator-based super-fidelity: A warm-start approach for accelerating steady-state simulations , author=. Journal of Computational Physics , volume=. 2025 , publisher=

    2025

  35. [35]

    International Journal of Mechanical Sciences , pages=

    A pretraining-finetuning computational framework for material homogenization , author=. International Journal of Mechanical Sciences , pages=. 2026 , publisher=

    2026

  36. [36]

    arXiv preprint arXiv:2601.03086 , year=

    Pretrain Finite Element Method: A Pretraining and Warm-start Framework for PDEs via Physics-Informed Neural Operators , author=. arXiv preprint arXiv:2601.03086 , year=

    arXiv

  37. [37]

    arXiv preprint arXiv:2511.06802 , year=

    Neural-initialized newton: Accelerating nonlinear finite elements via operator learning , author=. arXiv preprint arXiv:2511.06802 , year=

    arXiv

  38. [38]

    arXiv preprint arXiv:2509.21751 , year=

    Reparameterizing 4DVAR with neural fields , author=. arXiv preprint arXiv:2509.21751 , year=

    Pith/arXiv arXiv

  39. [39]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

  40. [40]

    Computer Methods in Applied Mechanics and Engineering , volume=

    A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2022 , publisher=

    2022

  41. [41]

    arXiv preprint arXiv:2511.08811 , year=

    A Neural-Operator Preconditioned Newton Method for Accelerated Nonlinear Solvers , author=. arXiv preprint arXiv:2511.08811 , year=

    arXiv

  42. [42]

    Advances in neural information processing systems , volume=

    Newton informed neural operator for solving nonlinear partial differential equations , author=. Advances in neural information processing systems , volume=

  43. [43]

    Computer Physics Communications , volume=

    High performance Python for direct numerical simulations of turbulent flows , author=. Computer Physics Communications , volume=. 2016 , publisher=

    2016

  44. [44]

    II: Modelling the forecast error covariance statistics , author=

    A review of forecast error covariance statistics in atmospheric variational data assimilation. II: Modelling the forecast error covariance statistics , author=. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography , volume=. 2008 , publisher=

    2008

  45. [45]

    I: Characteristics and measurements of forecast error covariances , author=

    A review of forecast error covariance statistics in atmospheric variational data assimilation. I: Characteristics and measurements of forecast error covariances , author=. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography , volume=. 2008 , publisher=

    2008

  46. [46]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Optimizing the optimizer for physics-informed neural networks and Kolmogorov-Arnold networks , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2025 , publisher=

    2025

  47. [47]

    State estimation in minimal turbulent channel flow: A comparative study of

    Du, Yifan and Wang, Mengze and Zaki, Tamer A , journal=. State estimation in minimal turbulent channel flow: A comparative study of. 2023 , publisher=

    2023

  48. [48]

    Nature methods , volume=

    SciPy 1.0: fundamental algorithms for scientific computing in Python , author=. Nature methods , volume=. 2020 , publisher=

    2020

  49. [49]

    zero-shot

    “zero-shot” super-resolution using deep internal learning , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  50. [50]

    Neurocomputing , volume=

    ZS-SRT: An efficient zero-shot super-resolution training method for Neural Radiance Fields , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  51. [51]

    The Twelfth International Conference on Learning Representations , year=

    Neural Spectral Methods: Self-supervised learning in the spectral domain , author=. The Twelfth International Conference on Learning Representations , year=

  52. [52]

    Part I: Technical formulation and preliminary test , author=

    An ensemble-based four-dimensional variational data assimilation scheme. Part I: Technical formulation and preliminary test , author=. Monthly Weather Review , volume=

  53. [53]

    Tellus A: Dynamic Meteorology and Oceanography , volume=

    Variational algorithms for analysis and assimilation of meteorological observations: theoretical aspects , author=. Tellus A: Dynamic Meteorology and Oceanography , volume=. 1986 , publisher=

    1986

  54. [54]

    Nonlinear Processes in Geophysics , volume=

    Nonlinear effects in 4D-Var , author=. Nonlinear Processes in Geophysics , volume=. 2018 , publisher=

    2018

  55. [55]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Multiple-truncation incremental approach for four-dimensional variational data assimilation , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 1998 , publisher=

    1998

  56. [56]

    Array programming with

    Harris, Charles R and Millman, K Jarrod and Van Der Walt, St. Array programming with. nature , volume=. 2020 , publisher=

    2020

  57. [57]

    Advances in neural information processing systems , volume=

    Characterizing possible failure modes in physics-informed neural networks , author=. Advances in neural information processing systems , volume=

  58. [58]

    Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=

    Understanding the difficulty of training deep feedforward neural networks , author=. Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=. 2010 , organization=

    2010

  59. [59]

    Nature , pages=

    Controlling diverse robots by inferring Jacobian fields with deep networks , author=. Nature , pages=. 2025 , publisher=

    2025

  60. [60]

    EGU General Assembly Conference Abstracts , pages=

    Coupling the data-driven weather forecasting model with 4d variational assimilation , author=. EGU General Assembly Conference Abstracts , pages=

  61. [61]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    DiffDA: a diffusion model for weather-scale data assimilation , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  62. [62]

    Advances in neural information processing systems , volume = 33, pages =

    Fourier features let networks learn high frequency functions in low dimensional domains , author =. Advances in neural information processing systems , volume = 33, pages =

  63. [63]

    Bioinformatics , publisher =

    JAX-RNAfold: scalable differentiable folding , author =. Bioinformatics , publisher =

  64. [64]

    2009 IEEE 12th international conference on computer vision , pages =

    What is the best multi-stage architecture for object recognition? , author =. 2009 IEEE 12th international conference on computer vision , pages =

    2009

  65. [65]

    , author =

    3D Gaussian splatting for real-time radiance field rendering. , author =. ACM Trans. Graph. , volume = 42, number = 4, pages =

  66. [66]

    Neural computation , publisher =

    Backpropagation applied to handwritten zip code recognition , author =. Neural computation , publisher =

  67. [67]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

    Deep image prior , author =. Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

  68. [68]

    SIAM Journal on Scientific Computing , publisher =

    Introducing time parallelization within data assimilation , author =. SIAM Journal on Scientific Computing , publisher =

  69. [69]

    Computer Methods in Applied Mechanics and Engineering , publisher =

    Forward and inverse simulation of pseudo-two-dimensional model of lithium-ion batteries using neural networks , author =. Computer Methods in Applied Mechanics and Engineering , publisher =

  70. [70]

    International Conference on Learning Representations , url =

    Decoupled Weight Decay Regularization , author =. International Conference on Learning Representations , url =

  71. [71]

    Blending Neural Operators and Relaxation Methods in PDE Numerical Solvers , author =

  72. [72]

    International conference on machine learning , pages =

    On the spectral bias of neural networks , author =. International conference on machine learning , pages =

  73. [73]

    Namgyu Kang and Jaemin Oh and Youngjoon Hong and Eunbyung Park , year = 2025, booktitle =

    2025

  74. [74]

    Atmospheric data analysis , author =

  75. [75]

    Advances in Neural Information Processing Systems , volume = 34, pages =

    Nerv: Neural representations for videos , author =. Advances in Neural Information Processing Systems , volume = 34, pages =

  76. [76]

    Emilien Dupont and Adam Golinski and Milad Alizadeh and Yee Whye Teh and Arnaud Doucet , year = 2021, booktitle =

    2021

  77. [77]

    Tellus A: Dynamic Meteorology and Oceanography , publisher =

    Incremental 4D-Var convergence study , author =. Tellus A: Dynamic Meteorology and Oceanography , publisher =

  78. [78]

    An easily implemented task-based parallel scheme for the

    Yin, Z and Clercx, HJH and Montgomery, DC , year = 2004, journal =. An easily implemented task-based parallel scheme for the

    2004

  79. [79]

    User guide to ECMWF forecast products , author =

  80. [80]

    Spectral methods: evolution to complex geometries and applications to fluid dynamics , author =

Showing first 80 references.